[
  {
    "benchmark": "swe-bench",
    "score": 34.2,
    "metric": "accuracy",
    "cost_per_instance": 0.24,
    "average_runtime": 1344.0,
    "full_archive": "https://results.eval.all-hands.dev/eval-21370501716-nemotron-3_litellm_proxy-openai-NVIDIA-Nemotron-3-Nano-30B-A3B-FP8_26-01-27-05-06.tar.gz",
    "tags": [
      "swe-bench"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-27T20:00:30.624153+00:00"
  },
  {
    "benchmark": "gaia",
    "score": 8.5,
    "metric": "accuracy",
    "cost_per_instance": 0.04,
    "average_runtime": 794.0,
    "full_archive": "https://results.eval.all-hands.dev/gaia/litellm_proxy-openai-NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/21412718442/results.tar.gz",
    "tags": [
      "gaia"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-27T20:00:30.624153+00:00"
  },
  {
    "benchmark": "commit0",
    "score": 6.2,
    "metric": "accuracy",
    "cost_per_instance": 0.03,
    "average_runtime": 21.0,
    "full_archive": "https://results.eval.all-hands.dev/commit0/litellm_proxy-openai-NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/25079507282/results.tar.gz",
    "tags": [
      "commit0"
    ],
    "agent_version": "v1.17.0",
    "submission_time": "2026-04-28T22:20:32+00:00",
    "eval_visualization_page": "https://laminar.sh/shared/evals/fb317122-02be-4878-a1be-7465af2fb5e7"
  },
  {
    "benchmark": "swe-bench-multimodal",
    "score": 20.9,
    "metric": "solveable_accuracy",
    "cost_per_instance": 0.4,
    "average_runtime": 1179.0,
    "full_archive": "https://results.eval.all-hands.dev/21402472780-nemotron-3_litellm_proxy-openai-NVIDIA-Nemotron-3-Nano-30B-A3B-FP8_26-01-27-17-53.tar.gz",
    "tags": [
      "swe-bench-multimodal"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-27T20:00:30.624153+00:00",
    "component_scores": {
      "solveable_accuracy": 20.9,
      "unsolveable_accuracy": 6.1,
      "combined_accuracy": 16.0,
      "solveable_resolved": 14,
      "solveable_total": 67,
      "unsolveable_resolved": 2,
      "unsolveable_total": 33
    }
  },
  {
    "benchmark": "swt-bench",
    "score": 7.6,
    "metric": "accuracy",
    "cost_per_instance": 0.19,
    "average_runtime": 896.3,
    "full_archive": "https://results.eval.all-hands.dev/swtbench/litellm_proxy-openai-NVIDIA-Nemotron-3-Nano-30B-A3B-FP8/21413263674/results.tar.gz",
    "tags": [
      "swt-bench"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-27T20:00:30.624153+00:00"
  }
]